NVIDIA Nemotron 3.5 Super VL and Bolt: The Launch Claims, Checked
NVIDIA shipped a 120B open vision-language agent model and two smaller Bolt models. Here is what is confirmed, what is a vendor claim and what is missing.
Jump to 9 sections
NVIDIA launched Nemotron 3.5 Super VL on October 15, 2026, a 120-billion-parameter open model with 12 billion active parameters for vision-language agents. It also released Bolt 74B-A9B, a 2-bit model built to fit in 24GB of memory. NVIDIA claims five times faster task completion than the nearest rival in its class, a claim we could not verify.
Model Drop covers new models, platforms, features and tools as they ship. This brief is for developers and teams deciding whether to try the Nemotron 3.5 family. It lists what NVIDIA confirmed, the one independent score available, and the details the launch left out.
What Did NVIDIA Release?
NVIDIA released three models in the Nemotron 3.5 family. Interesting Engineering's report says Super VL has 120 billion total parameters, with 12 billion active per token. It handles reasoning, tool use and understanding of text, images, documents and video.
The two Bolt models are smaller. According to Let's Data Science, Bolt 74B-A9B has 74 billion total parameters and 9 billion active. It uses native 2-bit quantization and is designed to fit in 24GB of memory. Bolt E4B is a 4GB model aimed at fast conversational and on-device agents, scheduled for October 23.
Super VL and Bolt 74B-A9B are available now on Hugging Face, OpenRouter and ModelScope, and through NVIDIA NIM. NVIDIA released open weights with supporting training materials.
How Was Super VL Trained?
NVIDIA says Super VL saw about 26 trillion tokens of pretraining, about 3.8 million post-training samples and reinforcement learning across dozens of environments. Those figures come from the company's announcement. They describe scale but say little about data quality or mix.
The model can be adapted with NVIDIA NeMo and Python-based pipelines, and teams can train it further on proprietary data. That matters for enterprises that want to specialize a model on internal documents.
Our guide to mixture-of-experts versus dense models explains why a 120B model with 12B active parameters can run faster than a dense model of similar total size. Only the active experts compute for each token, but all weights still need to fit in memory.
Does the Five-Times-Faster Claim Hold Up?
We cannot confirm it. NVIDIA says Super VL completes tasks five times faster than the nearest competitor in its class. Coverage gives no test setup, hardware, workload or competitor name, so the number cannot be reproduced.
The accuracy claim has the same problem. NVIDIA says the model approaches its largest reasoning model's accuracy with roughly one-fifth as many parameters. No benchmark table accompanies that statement in the reports we reviewed.
Speed claims depend on batch size, precision, serving software and prompt length. A fair comparison needs the same hardware and workload. Our primer on how to read a model launch announcement lists the questions to ask before trusting a headline multiplier.
What Do Independent Scores Say?
One independent number exists. Let's Data Science reported that Artificial Analysis gave Super VL an Intelligence Index score of 26.5 and Bolt 74B-A9B a score of 21.4. Artificial Analysis also lists Super VL at 120 billion total parameters with 12 billion active.
Those scores place Bolt about five points below Super VL, a reasonable gap given its smaller active size and 2-bit weights. An index score is a composite, so it hides strengths and weaknesses on specific tasks. It also does not tell you how the model performs on your data.
| Model | Total / active parameters | Artificial Analysis score | Notes |
|---|---|---|---|
| Nemotron 3.5 Super VL | 120B / 12B | 26.5 | Vision-language agent model |
| Nemotron 3.5 Bolt 74B-A9B | 74B / 9B | 21.4 | 2-bit, fits 24GB |
| Nemotron 3.5 Bolt E4B | 4GB model | Not reported | Scheduled October 23 |
Who Is Already Using It?
NVIDIA named several early adopters. Scale AI built a vendor credit finance agent for Skydance and an M&A due diligence agent for EY. AfterQuery is using it for credit agreement analysis for Hebbia.
AI21 and Alice are testing cybersecurity attack resilience, and Mercor is evaluating coding performance on APEX-SWE. These are company-selected examples. They show intended use cases, not results.
The pattern is clear: long documents, financial analysis and multi-step workflows. That fits a vision-language model that can read charts, tables and scans along with text.
Can You Run Bolt Locally?
Bolt 74B-A9B is designed to fit in 24GB of memory using native 2-bit quantization. That puts it in range of high-end consumer GPUs and workstations. NVIDIA positions it for local research, coding and multi-step work.
Native quantization means the model was trained at low precision, which usually costs less quality than compressing a model afterward. The Artificial Analysis score of 21.4 suggests a real gap from Super VL, though. Test it on your own tasks before relying on it.
If you plan to host it yourself, see our comparison of local AI inference servers. Check that your serving stack supports the 2-bit format before you download weights.
What Is Missing From the Launch?
Several details that developers need were not in the coverage. The license name, context window length, pricing and detailed benchmark tables were not stated. Architecture details beyond parameter counts were thin as well.
Those gaps matter for adoption. A license decides whether you can use the model commercially, and context length decides whether it fits your documents. Check the model card on Hugging Face for both before you build on it.
At Model Drop, our editors read open-weight launches against a short checklist: license, context, hardware, third-party scores and pricing. Nemotron 3.5 clears the availability and hardware items but leaves the rest to verify.
How Does It Fit the Open-Weight Landscape?
It adds a large, agent-focused vision-language option from a hardware vendor with strong incentives to make models run well on its chips. Our explainer on open-weight versus closed models covers the tradeoffs of hosting your own.
The main appeal is control. You can run the model in your own environment, fine-tune it on private data and avoid per-token fees from a hosted provider. The main cost is the work of serving it.
Expect a flurry of community tests over the next week. Independent speed and accuracy numbers will arrive, and they may differ from NVIDIA's claims.
What Should You Do Now?
Try it if you build document-heavy agents and have the hardware. Run a small test on your own files and measure accuracy, latency and cost against what you use now. Ignore headline multipliers until you have your own numbers.
If you only need a hosted API, check OpenRouter's listing for pricing and rate limits. If you plan to self-host, confirm the license first. Keep your current model as a fallback during evaluation.
Next step: pull the model card from Hugging Face, read the license and context length, and run ten real tasks through the model this week.
